diff --git a/megatron/core/distributed/fsdp/src/megatron_fsdp/megatron_fsdp.py b/megatron/core/distributed/fsdp/src/megatron_fsdp/megatron_fsdp.py
index 6987729ba8f..f79a64c2126 100644
--- a/megatron/core/distributed/fsdp/src/megatron_fsdp/megatron_fsdp.py
+++ b/megatron/core/distributed/fsdp/src/megatron_fsdp/megatron_fsdp.py
@@ -1164,7 +1164,7 @@ def reset_mixed_precision_policy(self, mixed_precision_policy: MixedPrecisionPol
         self.mp_policy = mp_policy_reset
         self.param_and_grad_buffer.mp_policy = mp_policy_reset
 
-    def start_param_sync(self, *unused, force_sync: bool = False, force_dispatch: bool = False):
+    def start_param_sync(self, *unused, force_sync: bool = False, force_dispatch: bool = False, sync_and_return: bool = False):
         """
         Initiates param sync (all-gather) communication operations for all model parameters.
 
@@ -1177,6 +1177,10 @@ def start_param_sync(self, *unused, force_sync: bool = False, force_dispatch: bo
                 other settings.
             force_dispatch (bool, optional): force dispatch regardless of other settings.
         """
+        if sync_and_return:
+            self.synchronize_param_gather()
+            return
+
         self._replace_param_with_raw_if_needed()
 
         if not force_sync and self.ddp_config.overlap_param_gather:
diff --git a/megatron/core/pipeline_parallel/schedules.py b/megatron/core/pipeline_parallel/schedules.py
index 24ec25e5150..a0b9a7cef3d 100644
--- a/megatron/core/pipeline_parallel/schedules.py
+++ b/megatron/core/pipeline_parallel/schedules.py
@@ -734,6 +734,15 @@ def forward_backward_no_pipelining(
 
     if getattr(config, 'fine_grained_activation_offloading', False):
         off_interface.reset()
+    # Reset all_gather_pipeline bucket status before next validation iteration
+    if forward_only:
+        for model_chunk in [model]:
+            if (
+                model_chunk.ddp_config.overlap_param_gather
+                and model_chunk.ddp_config.use_megatron_fsdp
+            ):
+                model_chunk.start_param_sync(sync_and_return=True)
+
 
     if config.timers is not None:
         config.timers('forward-backward').stop()
@@ -1954,6 +1963,15 @@ def pp_post_backward(input_tensor_grad, vp_stage=None):
 
     if getattr(config, 'fine_grained_activation_offloading', False):
         off_interface.reset()
+    # Reset all_gather_pipeline bucket status before next validation iteration
+    if forward_only:
+        for model_chunk in model:
+            if (
+                model_chunk.ddp_config.overlap_param_gather
+                and model_chunk.ddp_config.use_megatron_fsdp
+            ):
+                model_chunk.start_param_sync(sync_and_return=True)
+
     # Restore config.grad_sync_func and config.param_sync_func.
     if forward_only:
         config.grad_sync_func, config.param_sync_func = grad_sync_func, param_sync_func
@@ -2361,6 +2379,14 @@ def enable_grad_sync():
 
     if getattr(config, 'fine_grained_activation_offloading', False):
         off_interface.reset()
+    # Reset all_gather_pipeline bucket status before next validation iteration
+    if forward_only:
+        for model_chunk in [model]:
+            if (
+                model_chunk.ddp_config.overlap_param_gather
+                and model_chunk.ddp_config.use_megatron_fsdp
+            ):
+                model_chunk.start_param_sync(sync_and_return=True)
 
     if config.timers is not None:
         config.timers('forward-backward').stop()